AI Product & Data · Builder · Writer

I build things at the intersection of AI systems and human behavior, turning data-driven insights into products people actually use. Currently seeking opportunities in AI Product / Data / Agent Development around the Yangtze River Delta (Shanghai, Hangzhou, Suzhou, Nanjing, Wuxi).

My core interest: how people behave inside systems, why they drift, and how to design tools that help them steer instead of drift.

• 🎓 M.S. in Business Analytics & AI from Johns Hopkins University (Carey), graduated Sep 2026
• 🌐 Personal site with a built-in AI Agent console: gaolihong.com
• ✍️ 85+ Chinese essays on culture, travel, and AI, 310,000+ characters
• 🔧 Stack: Python / SQL / TypeScript / Next.js / SwiftUI / LLM Agents / RAG / MCP
• 📍 Shanghai-based, Yangtze River Delta preferred
• 🧱 Building: cyber-ququ debate platform, AI quant trading system (simulation), Understanding AI from Scratch and Seeing the World essay series
• 🤖 Building AI-powered automation for my job search: multi-agent orchestration, semantic matching, weekly pipeline. Private project, but the patterns show up in my public work.

Next.js 16 · TypeScript · Tailwind CSS 4 · Better-Auth · Agent Harness

• AI Agent Console: a full Harness Loop with Chat / Cowork modes, built-in tools, pluggable skills, and persistent memory, with an admin dashboard included.
• Content platform: 85+ Markdown blog posts with Shiki syntax highlighting and KaTeX math, migrated from Hexo.
• Infrastructure: Cloudflare CDN and Docker stack on Alibaba Cloud Singapore ECS.
• Extras: RSS, sitemap, email subscriptions, and a friend-link system, all manageable from the UI.

Python · NumPy · Next.js · Vercel Serverless

• 8 model providers in one interface: Gemini, OpenAI, DeepSeek, Kimi, Zhipu GLM, Xiaomi MiMo, Anthropic Claude, and Hugging Face.
• Smart retrieval: linearized 10-K table chunks, company scoping, and multi-query expansion, using pure NumPy cosine similarity with no vector DB dependency.
• Grounded answers: every reply cites [Company p.N] with source passages shown inline.
• Evaluation harness: LLM-judged benchmark across 5 SEC filings, with configurable system prompts and parameters.

JHU AI Essentials for Business final project. Live demo: rag-studio.gaolihong.com.

Swift · SwiftUI · SwiftData · Local-first

• Five modules in one app: capture, tasks, knowledge, finance, and lifestyle, all local-first with private sync.
• Native macOS experience: designed for the menu bar and system integration, not another web app in a wrapper.
• Homebrew distribution: available via brew install epiphany-leon/nexalife/nexalife.
• Design philosophy: Steer, don't drift. Built around the idea that a personal system should guide, not constrain.

Successor to my earlier LifeOS template, a web version I de-identified and open-sourced before going native.

Python · PyTorch · Reinforcement Learning · Jupyter

• Tabular Q-learning agent that coordinates dynamic pricing, discounting, bundling, and price-discipline choices over a 120-day retail horizon.
• 62.6% average daily profit lift compared to the fixed-price baseline.
• Exploration-exploitation balance tested across multiple reward designs and state-space abstractions.
• Final project for BU.520.750 (Reinforcement Learning) at JHU Carey, Spring 2026.

TypeScript · Streamlit · Python · AI Report Generation

• Configurable dashboard: bring your own data and database, zero-config demo mode, pluggable AI risk reports.
• Data product thinking: modular data sources, customizable risk scoring, and templated report generation.
• Framework-first design: built as a reusable template rather than a one-off analysis.

Languages & Data

AI & Machine Learning

Web & Product

Native & Apple

Cloud & DevOps

• 🌐 Website & Blog: gaolihong.com
• 📧 Email: lihong.gao2002@gmail.com
• 💼 Portfolio: Lihong Gao